Enterprise Integration Patterns (EIP) identifies several patterns that have the capability to buffer messages. For example, an Aggregator buffers messages until they can be released and a QueueChannel buffers messages until consumers explicitly receive those messages from that channel. Because of the failures that can occur at any point within your message flow, EIP components that buffer messages also introduce a point where messages could be lost. To mitigate the risk of losing Messages, EIP defines the Message Store pattern which allows EIP components to store Messages typically in some type of persistent store (e.g. RDBMS). Spring Integration provides support for the Message Store pattern by a) defining a org.springframework.integration.store.MessageStore strategy interface, b) providing several implementations of this interface, and c) exposing a message-store attribute on all components that have the capability to buffer messages so that you can inject any instance that implements the MessageStore interface. Details on how to configure a specific Message Store implementation and/or how to inject a MessageStore implementation into a specific buffering component are described throughout the manual (see the specific component, such as QueueChannel, Aggregator, Resequencer etc.), but here are a couple of samples to give you an idea: QueueChannel ]]> Aggregator ]]> By default Messages are stored in-memory using org.springframework.integration.store.SimpleMessageStore, an implementation of MessageStore. That might be fine for development or simple low-volume environments where the potential loss of non-persistent messages is not a concern. However, the typical production application will need a more robust option, not only to mitigate the risk of message loss but also to avoid potential out-of-memory errors. Therefore, we also provide MessageStore implementations for a variety of data-stores. Below is a complete list of supported implementations: - uses RDBMS to store Messages - uses Redis key/value datastore to store Messages - uses MongoDB document store to store Messages - uses Gemfire distributed cache to store Messages However be aware of some limitations while using persistent implementations of the MessageStore. The Message data (payload and headers) is serialized and deserialized using different serialization strategies depending on the implementation of the MessageStore. For example, when using JdbcMessageStore, only Serializable data is persisted by default. In this case non-Serializable headers are removed before serialization occurs. Also be aware of the protocol specific headers that are injected by transport adapters (e.g., FTP, HTTP, JMS etc.). For example, <http:inbound-channel-adapter/> maps HTTP-headers into Message Headers and one of them is an ArrayList of non-Serializable org.springframework.http.MediaType instances. However you are able to inject your own implementation of the Serializer and/or Deserializer strategy interfaces into some MessageStore implementations (such as JdbcMessageStore) to change the behaviour of serialization and deserialization. Special attention must be paid to the headers that represent certain types of data. For example, if one of the headers contains an instance of some Spring Bean, upon deserialization you may end up with a different instance of that bean, which directly affects some of the implicit headers created by the framework (e.g., REPLY_CHANNEL or ERROR_CHANNEL). Currently they are not serializable, but even if they were the deserialized channel would not represent the expected instance. As a workaround we suggest to remove bean-ref headers via a <header-filter/> before sending a message to an endpoint backed by a persistent MessageStore. Also, we recommend using channel names instead of channel instances when setting those types of headers, thus allowing it to be resolved in real time by the ChannelResolver. Also avoid configuration of a message-flow like this: gateway -> queue-channel (backed by a persistent Message Store) -> service-activator That gateway creates a Temporary Reply Channel in the background, and it will be lost by the time the service-activator's poller reads from the queue, because it has been deserialized by another thread on the sending side. Nevertheless we are constantly thinking about potential improvements to the framework, such as a way to provide some robust default serialization strategy for messages in these cases.